Towards virtual cells - the need for actionable, robust perturbation models

Computational cell biology is evolving from descriptive atlases to predictive, actionable models — from mapping what cells are to simulating what they do. In this talk, I will outline progress toward virtual cells, focusing on machine learning approaches that enable robust perturbation modeling.

Among others, I will present scConcept, a framework that differs from large-scale foundation models such as Geneformer by learning in the latent space through control-based objectives. Rather than passively embedding cellular states, scConcept explicitly models transitions between them, capturing how cells move through gene-expression space in response to context and perturbation.

Building on this foundation, I will introduce CellFlow, a generative perturbation model that predicts how interventions — such as drugs, cytokines, or gene edits — reshape cellular phenotypes. By learning causal directions of change, CellFlow enables in silico experimentation and virtual screening of differentiation protocols.

Together, these developments point toward virtual cells: computational counterparts capable of robustly predicting and designing biological behavior.